Intelligent cement plant dust identification method based on image processing

Abstract

The invention relates to the technical field of image processing, and discloses a cement plant dust intelligent identification method based on image processing. The method comprises the following steps: extracting a crusher dust characteristic capsule group through a capsule network architecture; establishing connection with a conveyer belt dust capsule layer by adopting a dynamic routing algorithm to obtain an equipment dust coupling coefficient matrix; reconstructing the coupling coefficient matrix space into a stock bin dust routing intensity mapping graph; quantitative processing is achieved through concentration index calculation, and production line dust concentration grade evaluation is obtained; and performing multi-source separation by using a connected domain analysis algorithm, and identifying a dust source associated with the cement equipment. According to the invention, the capsule network and routing path visualization technology is introduced into the field of industrial dust monitoring for the first time, so that the technical levels of dust source accurate positioning, concentration quantitative evaluation and intelligent decision support are improved.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to an intelligent dust recognition method for cement plants based on image processing. Background Art

[0002] Existing cement plant dust monitoring technologies primarily rely on traditional image processing methods and convolutional neural networks (CNNs) for dust identification. These methods deploy surveillance cameras in key areas of the cement plant to capture real-time images of the production site. They then employ image preprocessing, feature extraction, and pattern recognition techniques to detect and identify dust areas within these images. Traditional methods typically employ algorithms such as threshold segmentation, edge detection, and region growing to separate dust areas from the background. These methods then combine machine learning algorithms such as support vector machines and decision trees to establish dust identification models, achieving a degree of automated dust detection.

[0003] However, existing technologies suffer from numerous limitations and technical deficiencies. The first is poor environmental adaptability. Cement plant production environments are complex and highly variable, and variations in lighting conditions, weather conditions, dust color, and background can significantly affect recognition accuracy. In particular, the maximum pooling operation in traditional CNN methods results in loss of positional information, making it impossible to accurately locate dust sources. Second, quantitative assessment capabilities are insufficient. Existing methods lack effective calibration methods and standard datasets, and are primarily limited to qualitative identification, making it difficult to accurately determine specific concentration values. Furthermore, existing technologies often exhibit "black box" characteristics and lack interpretability, failing to provide a reliable basis for decision-making.

[0004] The fundamental problem with existing dust identification technology lies in its inability to simultaneously address the three key technical challenges of precise positioning, quantitative assessment, and interpretability. Traditional methods, unable to maintain accurate location information, are unable to accurately locate multi-source dust and identify associated equipment. The lack of quantitative assessment capabilities prevents existing technologies from reliably mapping image features to specific concentration values. The lack of interpretability further limits the technology's practical application and credibility. These interrelated and mutually constraining issues create technical bottlenecks that are difficult to overcome with existing technologies. A new intelligent dust identification method for cement plants is urgently needed that can simultaneously address precise positioning, quantitative assessment, and decision support. Summary of the Invention

[0005] This application provides an image processing-based intelligent dust identification method for cement plants. This method addresses the technical issues of existing image processing-based methods for cement plant dust identification, such as the inability to accurately locate dust sources, the lack of quantitative concentration assessment capabilities, and insufficient interpretability for decision support. By introducing capsule networks and routing path visualization technology to the field of industrial dust monitoring for the first time, this approach advances the state-of-the-art in dust source location, quantitative concentration assessment, and intelligent decision support.

[0006] The present application provides a method for intelligently identifying dust in a cement plant based on image processing, which includes: Step S1: feature extraction and processing of cement plant monitoring images are performed through a capsule network architecture to obtain a crusher dust feature capsule group; Step S2: performing routing connection processing on the crusher dust feature capsule group and the conveyor belt dust capsule layer according to a dynamic routing algorithm to obtain an equipment dust coupling coefficient matrix; Step S3: spatially reconstructing the equipment dust coupling coefficient matrix to obtain a silo dust routing intensity map; Step S4: performing concentration quantification processing on the silo dust route intensity map by calculating the cement plant dust concentration index to obtain a production line dust concentration level assessment result; In step S5, a multi-source separation process is performed on the silo dust routing intensity map according to a connected domain analysis algorithm to obtain a cement equipment-associated dust source identification result.

[0007] Optionally, the step S1 includes: Perform histogram equalization and gamma correction on the cement plant monitoring image to obtain a standardized pre-processed image; Inputting the standardized preprocessed image into the primary capsule layer of the capsule network architecture for convolution kernel feature extraction processing to obtain a crusher region capsule vector group; Performing nonlinear transformation processing on the crusher region capsule vector group based on the ReLU activation function to obtain an activated crusher feature vector; The activated crusher feature vector is subjected to length restriction processing according to a vector length normalization algorithm to obtain a crusher dust feature capsule group.

[0008] Optionally, the step S2 includes: The crusher dust feature capsule group is input as a primary capsule into the dynamic routing algorithm for initialization processing to obtain an initial coupling coefficient matrix; Iteratively updating the initial coupling coefficient matrix according to the consistency between the prediction vector and the output vector to obtain an updated coupling coefficient matrix; Performing weight normalization processing on the updated coupling coefficient matrix based on a softmax normalization function to obtain a normalized coupling weight matrix; Performing a weighted sum operation on the normalized coupling weight matrix and the conveyor belt dust capsule layer to obtain a capsule layer output vector; The coupling coefficient of the capsule layer output vector is determined according to the vector length calculation to obtain the equipment dust coupling coefficient matrix.

[0009] Optionally, the iterative updating of the initial coupling coefficient matrix according to the consistency between the prediction vector and the output vector to obtain an updated coupling coefficient matrix includes: Calculating the prediction vector between the crusher dust feature capsule group and the conveyor belt dust capsule layer to obtain an inter-capsule prediction vector matrix; Performing a dot product operation on the inter-capsule prediction vector matrix and the capsule layer output vector to obtain a consistency score matrix; Performing weighted accumulation processing on the consistency score matrix and the initial coupling coefficient matrix to obtain an accumulated coupling coefficient matrix; The accumulated coupling coefficient matrix is ​​updated by three rounds of iterative cycles to obtain an updated coupling coefficient matrix.

[0010] Optionally, the step S3 includes: Extracting the receptive field position information of each crusher dust feature capsule in the equipment dust coupling coefficient matrix to obtain a capsule space coordinate mapping table; Performing a two-dimensional spatial reconstruction process on the device dust coupling coefficient matrix according to the capsule space coordinate mapping table to obtain a discrete intensity distribution map; Performing continuous expansion processing on the discrete intensity distribution map based on a bilinear interpolation algorithm to obtain a continuous intensity distribution map; The continuous intensity distribution map is subjected to Gaussian filtering and smoothing processing to obtain a silo dust routing intensity map.

[0011] Optionally, performing two-dimensional spatial reconstruction processing on the device dust coupling coefficient matrix according to the capsule space coordinate mapping table to obtain a discrete intensity distribution map includes: Calculate the center coordinate position of each crusher dust feature capsule in the original image according to the capsule space coordinate mapping table to obtain a capsule center coordinate array; Performing spatial position matching processing on the coupling coefficient values ​​in the equipment dust coupling coefficient matrix according to the capsule center coordinate array to obtain a coordinate-intensity correspondence table; Based on the coordinate-intensity correspondence table, a two-dimensional array structure of the same size as the cement plant monitoring image is constructed to obtain an initialized intensity matrix; The intensity values ​​in the coordinate-intensity correspondence table are filled into the corresponding coordinate positions of the initialized intensity matrix to obtain a discrete intensity distribution map.

[0012] Optionally, the step S4 includes: Counting the total area and average intensity value of the area above the intensity threshold in the dust route intensity map of the silo to obtain dust area statistical parameters; Performing a product operation on the dust area statistical parameters and the calibration coefficient to obtain a cement plant dust concentration index; Inputting the cement plant dust concentration index into a multiple linear regression model for concentration numerical conversion processing to obtain a quantitative dust concentration value; The quantified dust concentration value is classified based on a preset concentration level classification standard to obtain a production line dust concentration level assessment result.

[0013] Optionally, the step S5 includes: Performing a connected domain analysis algorithm on the silo dust routing intensity map to obtain a connected domain set of independent dust regions; Calculating the geometric characteristic parameters of each connected domain in the set of connected domains of the independent dust regions to obtain a dust source geometric characteristic database; Inputting the dust source geometric feature database into a support vector machine classifier to perform equipment type identification processing to obtain a dust source equipment classification result; The dust source equipment classification results are subjected to position matching and association processing according to the cement plant equipment spatial layout information to obtain cement equipment associated dust source identification results.

[0014] Optionally, the calculating of the geometric characteristic parameters of each connected domain in the set of connected domains of the independent dust regions to obtain a dust source geometric characteristic database includes: Calculating the coordinates of the center of gravity and the size of the circumscribed rectangle of each connected domain in the set of connected domains of the independent dust regions to obtain characteristic parameters of the crusher dust source position; Calculate the aspect ratio and regional compactness value according to the crusher dust source position characteristic parameters to obtain the conveyor belt dust source shape characteristic parameters; Performing feature vector combination processing based on the shape characteristic parameters of the conveyor belt dust source and the area of ​​the connected domain to obtain a comprehensive feature vector of the silo dust source; The comprehensive feature vector of the silo dust source is processed according to the connected domain number to construct a feature database to obtain a dust source geometric feature database.

[0015] In the technical solution provided in this application, feature extraction and processing of cement plant monitoring images are performed through a capsule network architecture, which has significant technical advantages over traditional CNN methods. The capsule network can maintain the precise spatial position information of the dust area in the image, avoiding the problem of position information loss caused by the maximum pooling operation, and laying a solid foundation for subsequent precise positioning. The dynamic routing algorithm establishes an intelligent connection between the crusher dust feature capsule group and the conveyor belt dust capsule layer. Through multiple rounds of iterative adaptive adjustment of the connection weights, a strong connection is formed between capsules with similar features, realizing a hierarchical mapping from underlying visual features to high-level semantic concepts, and solving the technical problem that the traditional fixed connection mode cannot adapt to complex dust scenes. The spatial reconstruction processing of the equipment dust coupling coefficient matrix innovatively converts the one-dimensional coupling relationship into a two-dimensional spatial intensity distribution. Combined with the receptive field position information and the bilinear interpolation algorithm, a continuous and smooth silo dust routing intensity map is generated. This technological breakthrough enables the abstract internal network features to be intuitively mapped to the actual physical space position.

[0016] The cement plant dust concentration index calculation algorithm fully considers the physical characteristics and spatial distribution of dust diffusion. Using a multivariate linear regression model, it establishes a precise mapping from routing intensity to actual concentration, achieving a fundamental shift from qualitative identification to quantitative assessment. The connected domain analysis algorithm, when addressing multi-source dust separation, incorporates the geometric characteristics of dust sources and the spatial layout of equipment, enabling accurate identification of characteristic patterns of dust generated by different equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A schematic diagram of an embodiment of an intelligent identification method for cement plant dust based on image processing; Figure 2 This is a schematic diagram of the dust concentration level evaluation results of the production line in the embodiment of this application. DETAILED DESCRIPTION

[0019] An embodiment of the present application provides an intelligent identification method for cement plant dust based on image processing. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or apparatus.

[0020] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 An embodiment of a cement plant dust intelligent identification method based on image processing includes: Step S1: feature extraction and processing of cement plant monitoring images are performed through a capsule network architecture to obtain a crusher dust feature capsule group; Step S2: perform routing connection processing on the crusher dust feature capsule group and the conveyor belt dust capsule layer according to the dynamic routing algorithm to obtain the equipment dust coupling coefficient matrix; Step S3: spatially reconstruct the equipment dust coupling coefficient matrix to obtain a silo dust routing intensity map; Step S4: quantify the dust concentration of the silo dust route intensity map by calculating the cement plant dust concentration index to obtain the dust concentration level assessment result of the production line; In step S5, a multi-source separation process is performed on the silo dust routing intensity map according to the connected domain analysis algorithm to obtain the cement equipment associated dust source identification result.

[0021] It is understandable that the execution subject of this application can be a cement plant dust intelligent identification system based on image processing, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0022] Specifically, the capsule network architecture extracts features from cement plant monitoring images through a primary capsule layer. Histogram equalization is performed on the original 512×512 pixel cement plant monitoring image to enhance image contrast by redistributing pixel grayscale values. Gamma correction is then performed to adjust the image brightness curve. The resulting standardized preprocessed image is then fed into the capsule network architecture. The primary capsule layer uses convolutional kernels to extract features from the standardized preprocessed image, generating a crusher region capsule vector set, each containing eight dimensions of feature information. The ReLU activation function resets negative elements in the crusher region capsule vector set to zero, while retaining positive elements unchanged, forming an activated crusher feature vector. A vector length normalization algorithm calculates the Euclidean length of each activated crusher feature vector, limiting the vector length to between 0 and 1 to ensure a uniform value range for different capsule vectors, resulting in a crusher dust feature capsule set.

[0023] The dynamic routing algorithm establishes a connection between the crusher dust feature capsule group and the conveyor belt dust capsule layer. The algorithm uses the crusher dust feature capsule group as the primary capsule input and initializes the coupling coefficients between all capsules to zero, forming an initial coupling coefficient matrix. The prediction vector is calculated by multiplying the weight matrix by the crusher dust feature capsule group to obtain an inter-capsule prediction vector matrix. This matrix describes the expected value of the primary capsule for the higher-level capsule output. The dot product operation multiplies the inter-capsule prediction vector matrix by the actual capsule layer output vector element-by-element and sums them to generate a consistency score matrix, which quantifies the degree of match between the prediction vector and the actual output. The weighted accumulation process numerically accumulates the consistency score matrix and the initial coupling coefficient matrix according to the set weights to form an accumulated coupling coefficient matrix. The three-round iterative update process repeats the above prediction, scoring, and accumulation operations. In each iteration, the coupling coefficient is adaptively adjusted based on the degree of consistency. After three iterations, a converged updated coupling coefficient matrix is ​​obtained. The softmax normalization function performs exponential and normalization on the updated coupling coefficient matrix, ensuring that the coefficients in each row sum to 1, generating a normalized coupling weight matrix. A weighted summation operation multiplies the normalized coupling weight matrix by the conveyor dust capsule layer to obtain the capsule layer output vector. Vector length calculation uses a square root operation to determine the modulus of the capsule layer output vector, which serves as the basis for determining the coupling coefficient and forming the equipment dust coupling coefficient matrix.

[0024] The spatial reconstruction process converts the equipment dust coupling coefficient matrix into an intensity map with spatial location information. The receptive field location information extraction process calculates the spatial location of each crusher dust feature capsule in the original image based on the convolution parameters of the primary capsule layer. A convolution kernel stride of 2 means that the centers of adjacent capsules are separated by 2 pixels. This creates a capsule spatial coordinate mapping table. The two-dimensional spatial reconstruction process calculates the center coordinates of each capsule in the original image. The values ​​in the equipment dust coupling coefficient matrix are then aligned with the corresponding spatial coordinates to establish a coordinate-intensity correspondence table. The intensity matrix is ​​initialized to construct a two-dimensional array of the same size as the original cement plant monitoring image. The intensity values ​​from the coordinate-intensity correspondence table are then applied to the corresponding coordinates to form a discrete intensity distribution map. A bilinear interpolation algorithm fills in the blank spaces in the discrete intensity distribution map. The intensity values ​​of unknown points are estimated by calculating the weighted average of the four adjacent known points, generating a continuous intensity distribution map. The Gaussian filter smoothing process uses a Gaussian kernel function to convolve the continuous intensity distribution map, removing image noise and smoothing the intensity distribution, resulting in a silo dust routing intensity map.

[0025] The cement plant dust concentration index calculation converts the silo dust routing intensity map into a quantitative concentration assessment result. The statistical processing sets the intensity threshold at 60% of the map's maximum intensity value, traverses all pixels in the silo dust routing intensity map, identifies areas above the intensity threshold, and calculates their total area. The arithmetic mean of all pixel intensity values ​​within these areas is also calculated to obtain the dust area statistical parameters. The product operation mathematically multiplies the statistically obtained total area and average intensity value by a preset calibration coefficient. The calibration coefficient is obtained through regression analysis of synchronized measurement data from actual dust concentration detection equipment. The calculated result is the cement plant dust concentration index. The multivariate linear regression model uses the cement plant dust concentration index as the input variable. It combines multiple characteristic parameters such as the maximum intensity value of the map and the area of ​​the effective dust area, performs a linear combination calculation using pre-trained regression coefficients, and outputs a quantitative dust concentration value in milligrams per cubic meter. The grade classification process classifies the quantitative dust concentration values ​​according to the concentration range set by the environmental protection standards. 0-50 mg / m3 is classified as mild, 50-150 mg / m3 is classified as medium, 150-300 mg / m3 is classified as heavy, and 300 mg / m3 and above is classified as severe, thus obtaining the dust concentration grade assessment result of the production line.

[0026] The connected domain analysis algorithm performs multi-source dust separation and equipment association identification on the silo dust routing intensity map. The connected domain analysis process uses the eight-connected algorithm to traverse the silo dust routing intensity map and identify interconnected pixel areas with intensity values ​​higher than the threshold. Each connected area represents an independent dust source. After filtering tiny areas with an area less than 200 pixels, a connected domain set of independent dust areas is obtained. The calculation of geometric feature parameters includes four parts: center of gravity coordinate calculation, circumscribed rectangle size measurement, aspect ratio calculation, and area compactness calculation. The center of gravity coordinates are obtained by the weighted average of all pixel coordinates in the connected domain. The circumscribed rectangle is determined by finding the minimum and maximum coordinate values ​​of the connected domain. The aspect ratio is the ratio of the length of the circumscribed rectangle to the width. The area compactness is calculated by four times pi multiplied by the area divided by the square of the perimeter. These parameters constitute the dust source geometric feature database. The support vector machine classifier uses a database of dust source geometric features as input feature vectors and uses a pre-trained classification model to distinguish the dust characteristics of different equipment types. Crusher dust is typically highly compact, conveyor belt dust exhibits a strip-like distribution, and silo dust exhibits a diffuse distribution. The classifier outputs the dust source equipment classification results. Position matching and association processing coordinate-matches the dust source equipment classification results with pre-stored cement plant equipment spatial layout information. An association relationship is established based on the principle of minimizing the distance between the dust source's center of gravity coordinates and the equipment location, resulting in cement equipment-associated dust source identification results.

[0027] In a specific embodiment, step S1 includes: Perform histogram equalization and gamma correction on the cement plant monitoring image to obtain a standardized pre-processed image; The standardized pre-processed image is input into the primary capsule layer of the capsule network architecture for convolution kernel feature extraction to obtain the crusher region capsule vector group; Based on the ReLU activation function, the crusher region capsule vector group is subjected to nonlinear transformation processing to obtain the activated crusher feature vector; The activated crusher feature vector is length-restricted according to the vector length normalization algorithm to obtain the crusher dust feature capsule group.

[0028] Specifically, the histogram equalization processing of cement plant monitoring images solves the problem of uneven image contrast by remapping the pixel grayscale value distribution. The algorithm counts the number of pixels at each grayscale level from 0 to 255 in the original image, establishes a grayscale histogram, and then calculates the cumulative distribution function, that is, the total number of pixels at each grayscale level and all grayscale levels below it divided by the total number of pixels in the image. The original grayscale value is then mapped to a new grayscale range through the cumulative distribution function. The mapping formula is that the new grayscale value is equal to the cumulative distribution function value multiplied by 255. During the parameter substitution process, assuming that the total number of pixels in a cement plant monitoring image is 262,144, there are 8,000 pixels with a grayscale value of 120, and the cumulative number of pixels with grayscale values ​​from 0 to 120 is 150,000, then the cumulative distribution function value of the grayscale value of 120 is 150,000 divided by 262,144, which equals 0.572. The new grayscale value after mapping is 0.572 multiplied by 255, which equals 146. This mapping disperses the pixels originally concentrated in the low-grayscale area across the entire grayscale range. Gamma correction uses a power function transformation to adjust the image brightness characteristics. The transformation formula is that the output pixel value is equal to the input pixel value divided by 255 raised to the gamma power, then multiplied by 255. When the input pixel value is 100 and the gamma value is 0.8, the output pixel value is 100 divided by 255 raised to the 0.8 power, then multiplied by 255, which is 114. This nonlinear transformation can compensate for the uneven lighting in the dusty environment of the cement plant and produce a standardized preprocessed image.

[0029] The primary capsule layer of the capsule network architecture extracts features from a standardized preprocessed image through convolution. As the convolution kernel slides across the image, it performs element-by-element multiplication and summation with the pixels in the corresponding region. For this calculation, let W be the 9×9 convolution kernel weight matrix and X be the pixel matrix of the corresponding image region. The convolution output is the inner product of W and X. In this example, the pixel value matrix of a 9×9 image region is convolved with the convolution kernel weight matrix, and the 81 product results are summed to obtain the feature response value at that location. When the convolution kernel slides across a 512×512 image with a stride of 2, the output feature map is 252×252 in size. 32 different convolution kernels generate 32 feature maps. The primary capsule layer reorganizes these 32 feature maps into 8-dimensional vector form. The corresponding pixel values ​​of every 8 feature maps form an 8-dimensional capsule vector. Therefore, the 252×252 feature map area produces 252×252 8-dimensional capsule vectors, and a total of 63,504 capsule vectors are generated. Each vector contains 8 numerical elements. These vectors constitute the crusher area capsule vector group.

[0030] The ReLU activation function performs a nonlinear transformation on the crusher region capsule vector group. The function is defined as outputting an input when the input is greater than 0 and outputting 0 when the input is less than or equal to 0. The algorithm processes each dimension element of the capsule vector one by one. During the data calculation process, suppose an 8-dimensional capsule vector is [-0.3, 0.7, -0.1, 0.9, 0.2, -0.5, 0.4, 0.6]. The ReLU function processes each element separately. If -0.3 is less than 0, the output is 0, 0.7 is greater than 0, the output is 0.7, -0.1 is less than 0, the output is 0, 0.9 is greater than 0, the output is 0.9, 0.2 is greater than 0, the output is 0.2, -0.5 is less than 0, the output is 0, 0.4 is greater than 0, and the output is 0.6 is greater than 0. The processed vector becomes [0, 0.7, 0, 0.9, 0.2, 0, 0.4, 0.6]. This processing method suppresses negative noise while retaining positive feature signals. The 63504 capsule vectors processed by ReLU constitute the activated crusher feature vector.

[0031] The vector length normalization algorithm uses a compression function to limit the length of the activated crusher feature vector. The compression function formula is: the output vector is equal to the vector length squared divided by 1 plus the vector length squared, multiplied by the unit vector obtained by dividing the input vector by its length. The specific calculation process is to calculate the Euclidean length of the vector, which is the square root of the sum of the squares of its elements. For the vector [0, 0.7, 0, 0.9, 0.2, 0, 0.4, 0.6], the length is calculated as the sum of 0 squared plus 0.7 squared plus 0 squared plus 0.9 squared plus 0.2 squared plus 0 squared plus 0.4 squared plus 0.6 squared, that is, 0 plus 0.49 plus 0 plus 0.81 plus 0.04 plus 0 plus 0.16 plus 0.36, which equals 1.86. The square root is 1.364. The square of the vector length in the compression function is 1.86, 1 plus the square of the vector length is 2.86, and the compression factor is 1.86 divided by 2.86, which equals 0.65. The unit vector is obtained by dividing the original vector by its length to obtain [0, 0.513, 0, 0.66, 0.147, 0, 0.293, 0.44]. The output vector is [0, 0.333, 0, 0.429, 0.095, 0, 0.19, 0.286] by the compression factor 0.65 multiplied by the unit vector. This normalization process ensures that the length of all capsule vectors is limited to between 0 and 1, while keeping the vector direction information unchanged. The 63,504 capsule vectors after length normalization constitute the crusher dust feature capsule group.

[0032] In a specific embodiment, step S2 includes: The crusher dust feature capsule group is used as the primary capsule to input the dynamic routing algorithm for initialization processing to obtain the initial coupling coefficient matrix; The initial coupling coefficient matrix is ​​iteratively updated according to the consistency between the prediction vector and the output vector to obtain an updated coupling coefficient matrix; The updated coupling coefficient matrix is ​​weight normalized based on the softmax normalization function to obtain a normalized coupling weight matrix; Perform weighted sum operation on the normalized coupling weight matrix and the conveyor belt dust capsule layer to obtain the capsule layer output vector; The coupling coefficient of the capsule layer output vector is determined based on the vector length calculation to obtain the equipment dust coupling coefficient matrix.

[0033] Specifically, when the crusher dust feature capsule group is input as a primary capsule into the dynamic routing algorithm for initialization, the algorithm establishes a connection between each primary capsule and each capsule in the conveyor belt dust capsule layer, initializing all coupling coefficients to zero. Assuming the crusher dust feature capsule group contains 63,504 8-dimensional capsule vectors and the conveyor belt dust capsule layer contains 10 16-dimensional capsule vectors, it is necessary to establish 63,504 times 10, which equals 635,040 connection relationships. Each connection corresponds to a coupling coefficient. The initial coupling coefficient matrix is ​​a 63,504-row, 10-column zero matrix, where each element represents the weight of the corresponding primary capsule routing to the conveyor belt dust capsule. The dynamic routing algorithm adjusts these coupling coefficients through multiple rounds of iterations, forming strong connections between capsules with similar features and weak connections between unrelated capsules, thereby establishing a hierarchical feature representation relationship.

[0034] Computing the consistency between the prediction vector and the output vector is a core mechanism of the dynamic routing algorithm, used to measure the accuracy of the primary capsule's prediction of the higher-level capsule's output. During the prediction vector calculation process, each primary capsule generates a prediction value for the conveyor belt dust capsule through a weight matrix transformation. The weight matrix consists of 8 rows and 16 columns of trainable parameters, mapping the 8-dimensional primary capsule vector to a 16-dimensional prediction vector. The consistency calculation uses a vector dot product operation, multiplying the prediction vector element-wise with the actual conveyor belt dust capsule output vector and then summing the results to obtain a scalar consistency score. When the prediction vector and the output vector are in the same direction, the dot product value is positive and large, indicating an accurate prediction. When the directions are opposite, the dot product value is negative, indicating an incorrect prediction. The iterative update process adds the consistency score to the corresponding initial coupling coefficient. After three rounds of iteration, the coupling coefficients of the connections with accurate predictions gradually increase, while the coupling coefficients of the connections with incorrect predictions remain smaller, resulting in an updated coupling coefficient matrix that reflects the strength of the association between capsules.

[0035] The softmax normalization function performs weight normalization on the updated coupling coefficient matrix to ensure that the sum of all connection weights emanating from each primary capsule is 1. The softmax function performs an exponential operation on each row of the coupling coefficient matrix, converting the original coefficient value into a positive value, and then calculates the sum of the exponential values ​​of each row, and finally divides each exponential value by the sum of the row to obtain the normalized weight. In the specific calculation process, the row vector of the coupling coefficient corresponding to a primary capsule is [-0.2, 0.5, -0.1, 0.8, 0.3, -0.3, 0.1, 0.4, 0.6, -0.1]. After the exponential operation, it is [0.819, 1.649, 0.905, 2.226, 1.350, 0.741, 1.105, 1.492, 1.822, 0.905]. The total of the row is 13.014, and the normalized weight vector is [0.063, 0.127, 0.070, 0.171, 0.104, 0.057, 0.085, 0.115, 0.140, 0.070], the sum of the weights is equal to 1 and each weight is positive. This normalization process makes the coupling coefficient have a probability distribution characteristic, which represents the probability of the primary capsule routing to the dust capsules on different conveyor belts.

[0036] The normalized coupling weight matrix is ​​then weighted and summed with the conveyor dust capsule layer to calculate the output vector for each conveyor dust capsule. During this weighted summation, the output of each conveyor dust capsule is equal to the weighted average of the prediction vectors of all primary capsules, with the weights coming from the corresponding columns of the normalized coupling weight matrix. The output vector of the j-th conveyor dust capsule is calculated as the sum of the prediction vectors of all primary capsules i multiplied by the corresponding normalized weights. This means the output vector equals the sum of weight 1 multiplied by prediction vector 1 plus weight 2 multiplied by prediction vector 2, and so on, up to weight 63504 multiplied by prediction vector 63504. This weighted summation mechanism gives primary capsules that accurately predict the j-th conveyor dust capsule a greater influence weight, while primary capsules that incorrectly predict have a smaller influence weight, resulting in a more accurate capsule layer output vector.

[0037] Vector length calculation determines the coupling coefficient of the capsule layer's output vector and quantifies the capsule activation strength by calculating the modulus of the output vector. This calculation uses the Euclidean distance formula, which is the square root of the sum of the squares of the elements in each dimension of the vector. For example, for a 16-dimensional conveyor belt dust capsule output vector [0.3, -0.2, 0.7, 0.5, -0.1, 0.8, 0.4, -0.3, 0.6, 0.2, -0.4, 0.9, 0.1, 0.5, -0.2, 0.7], the vector length is calculated as the square root of the sum of the squares of 0.3 squared plus -0.2 squared plus 0.7 squared, resulting in a value of 1.856. In a capsule network, vector length represents the probability of detecting the corresponding entity. A length close to 1 indicates a high probability of detecting that type of dust, while a length close to 0 indicates that the dust type was not detected. The coupling coefficient determination process uses the calculated vector length as the activation value of the capsule, while maintaining the attribute parameters of the vector direction information encoding entity. The length values ​​of the 10 conveyor belt dust capsules constitute a row of the equipment dust coupling coefficient matrix. After processing multiple image samples, the equipment dust coupling coefficient matrix is ​​formed.

[0038] In a specific embodiment, the process of performing the iterative update process on the initial coupling coefficient matrix according to the consistency between the prediction vector and the output vector may specifically include the following steps: Calculate the prediction vector between the crusher dust feature capsule group and the conveyor belt dust capsule layer to obtain the inter-capsule prediction vector matrix; Perform dot product operations on the inter-capsule prediction vector matrix and the capsule layer output vector to obtain the consistency score matrix; Perform weighted accumulation processing on the consistency score matrix and the initial coupling coefficient matrix to obtain the accumulated coupling coefficient matrix; The accumulated coupling coefficient matrix is ​​updated by three rounds of iterative cycles to obtain an updated coupling coefficient matrix.

[0039] Specifically, the prediction vector calculation process between the crusher dust feature capsule group and the conveyor belt dust capsule layer establishes a mapping relationship between capsules through a trainable transformation matrix. Each 8-dimensional crusher dust feature capsule is transformed into a 16-dimensional prediction vector using an 8-row, 16-column weight matrix. Each element in the transformation matrix is ​​a parameter learned through training, reflecting the strength of the association between features of different dimensions. Matrix multiplication multiplies each dimension of the crusher dust feature capsule by the corresponding row of the weight matrix element-by-element, and the sum is calculated to obtain the value of each dimension of the prediction vector. In the specific calculation process, assume that the crusher dust feature capsule vector is [0.3, 0.7, 0.2, 0.8, 0.5, 0.1, 0.6, 0.4] and the first row of the weight matrix is ​​[0.2, -0.1, 0.3, 0.5, -0.2, 0.4, 0.1, 0.6]. The first dimension of the prediction vector is equal to the sum of 0.3 times 0.2 plus 0.7 times negative 0.1 plus 0.2 times 0.3, and so on, resulting in a total of 1.12. Each crusher dust feature capsule generates a prediction vector for each of the ten conveyor belt dust capsules using its own weight matrix. This results in a 16-dimensional inter-capsule prediction vector matrix with 63,504 rows and 10 columns. This matrix contains 635,040 16-dimensional prediction vectors, each representing the expected output of the corresponding crusher capsule for a specific conveyor belt dust capsule.

[0040] The prediction accuracy is calculated by performing a dot product operation on the inter-capsule prediction vector matrix and the capsule layer output vector. The dot product operation multiplies the corresponding dimension elements of the two vectors and then sums them to obtain a scalar result. During the dot product calculation, suppose a 16-dimensional prediction vector is [0.3, -0.2, 0.7, 0.5, -0.1, 0.8, 0.4, -0.3, 0.6, 0.2, -0.4, 0.9,0.1, 0.5, -0.2, 0.7], and the corresponding conveyor belt dust capsule output vector is [0.2, 0.1, 0.8, 0.4, 0.3, 0.7, 0.5, 0.2, 0.6, 0.3, 0.1, 0.8, 0.2, 0.4, 0.1, 0.6]. The dot product result is 0.3 multiplied by 0.2 plus negative 0.2 multiplied by 0.1 plus 0.7 multiplied by 0.8, and so on, the sum of 16 products is calculated to be 2.84. When the predicted and output vectors align in direction, the dot product is positive and large, indicating an accurate prediction. When they align in opposite directions, the dot product is negative, indicating an incorrect prediction. The 63,504 crusher dust feature capsules were dot-producted with the 10 conveyor belt dust capsules, resulting in a 63,504-row, 10-column consistency score matrix. Each element in the matrix quantifies the prediction accuracy of the corresponding capsule connection.

[0041] The consistency score matrix and the initial coupling coefficient matrix are weighted and accumulated to update the connection strength between capsules. The weighted accumulation algorithm adds the consistency score to the corresponding coupling coefficient according to the set weight. During the weighted accumulation process, the weight parameter controls the influence of the consistency score on the coupling coefficient in each iteration and is typically set to a small value between 0.1 and 0.3 to prevent system oscillation. Specifically, assuming an initial coupling coefficient of 0.0, a consistency score of 2.84 in the first iteration, and a weight of 0.2, the accumulated coupling coefficient equals 0.0 plus 2.84 times 0.2, which equals 0.568. Connections with accurate predictions have positive consistency scores, and the accumulated coupling coefficient increases. Connections with incorrect predictions have negative consistency scores, and the accumulated coupling coefficient decreases or even becomes negative. Weighted accumulation is performed on 63504 times 10 connections to produce the accumulated coupling coefficient matrix, which reflects the strength distribution of each capsule connection after one round of updates.

[0042] The three-round iterative cyclic update process gradually optimizes the coupling coefficient distribution by repeatedly executing prediction, scoring, and accumulation operations. Each round of iteration calculates a new conveyor belt dust capsule output vector based on the current coupling coefficient. The first round of iteration uses the initial all-zero coupling coefficient to calculate the output vector. The second round of iteration recalculates the output vector using the coupling coefficient updated in the first round. The third round of iteration uses the coupling coefficient updated in the second round to calculate the output vector again. The output vector of the conveyor belt dust capsule changes in each round of iteration, causing the consistency score to be recalculated, which in turn affects the update direction and amplitude of the coupling coefficient. After three rounds of iteration, the crusher capsules related to the conveyor belt dust characteristics form a stable strong connection with the corresponding conveyor belt dust capsules, and the coupling coefficient converges to a large positive value, while the unrelated capsule connections maintain a small coupling coefficient, resulting in a converged updated coupling coefficient matrix.

[0043] In a specific embodiment, step S3 includes: Extract the receptive field position information of each crusher dust feature capsule in the equipment dust coupling coefficient matrix to obtain the capsule space coordinate mapping table; The equipment dust coupling coefficient matrix is ​​reconstructed in two dimensions according to the capsule space coordinate mapping table to obtain a discrete intensity distribution map; The discrete intensity distribution map is continuously expanded based on the bilinear interpolation algorithm to obtain a continuous intensity distribution map; The continuous intensity distribution map is smoothed by Gaussian filtering to obtain the silo dust routing intensity map.

[0044] Specifically, the process of extracting the receptive field position information of each crusher dust feature capsule in the equipment dust coupling coefficient matrix calculates the spatial position of the capsule in the original image based on the convolutional layer parameters of the capsule network. The receptive field refers to the area of ​​the original image corresponding to each capsule, and its position is determined by the kernel size, stride, and padding parameters of the convolutional layer. The primary capsule layer uses a parameter configuration of a nine-by-nine convolution kernel and a stride of two. The center coordinates of the receptive field of each capsule are calculated using the formula, that is, the center horizontal coordinate is equal to the column index of the capsule in the feature map multiplied by the stride plus the convolution kernel radius, and the center vertical coordinate is equal to the row index of the capsule in the feature map multiplied by the stride plus the convolution kernel radius. In the specific calculation process, assuming that the position of a capsule in the 252×252 feature map is the 50th row and 80th column, then its center coordinates in the original 512×512 image are the horizontal coordinate 80 times 2 plus 4, which equals 164, and the vertical coordinate 50 times 2 plus 4, which equals 104. The 63,504 crusher dust feature capsules respectively calculate their receptive field center coordinates to form a capsule space coordinate mapping table containing horizontal and vertical coordinate pairs. This mapping table establishes the correspondence between the capsule index and the image space position.

[0045] The capsule spatial coordinate mapping table performs a two-dimensional spatial reconstruction of the device-dust coupling coefficient matrix, converting the one-dimensional capsule coupling coefficients into a two-dimensional intensity map with spatial distribution characteristics. The device-dust coupling coefficient matrix contains coupling coefficient values ​​corresponding to 63,504 capsules, each coefficient representing the strength of the connection between that capsule and a capsule of a specific dust type. The two-dimensional spatial reconstruction process creates a 512 by 512 zero-value matrix of the same size as the original image as the initial intensity matrix. The coupling coefficient values ​​of each capsule are then assigned to their corresponding spatial locations according to the capsule spatial coordinate mapping table. Specifically, assuming the coupling coefficient of the 1000th capsule is 0.75 and its corresponding spatial coordinates are 164 by 104, the value 0.75 is assigned to the 104th row and 164th column of the initial intensity matrix. Because the number of capsules is smaller than the number of image pixels, the reconstructed matrix contains a large number of blank locations, which remain zero, forming a sparsely distributed discrete intensity distribution map. This image reflects the discrete spatial distribution characteristics of the capsule coupling strength.

[0046] The bilinear interpolation algorithm continuously expands the discrete intensity distribution map to fill gaps and generate a smooth intensity distribution. Bilinear interpolation is an interpolation method based on a weighted average of the four nearest neighbor points. The algorithm determines four points with known intensity values ​​around the point to be interpolated and then calculates the interpolation result based on the distance weights. During the interpolation calculation, the coordinates of the point to be interpolated are x / y, and the coordinates of the four nearest neighbor known points are x / y, x / y, x / y, x / y, x / y, x / y, and x / y, respectively, with corresponding intensity values ​​of f11, f12, f21, and f22. Bilinear interpolation performs linear interpolation in the abscissa direction, calculating the interpolation results at y1 and y2, and then performing linear interpolation in the ordinate direction to obtain the result. The linear interpolation formula is: the interpolation result equals the left endpoint value multiplied by the right endpoint distance ratio plus the right endpoint value multiplied by the left endpoint distance ratio. After bilinear interpolation processing, the originally sparse discrete intensity distribution map becomes a dense continuous intensity distribution map, the blank areas are filled with reasonable intensity values, and the entire image presents a smooth intensity gradient feature.

[0047] The continuous intensity distribution map is smoothed using Gaussian filtering to eliminate noise and further enhance the image's continuity. Gaussian filtering is a linear smoothing filter based on the Gaussian function. It performs a local weighted average on the image through a convolution operation. The Gaussian kernel function uses the center pixel as the reference, with closer pixels receiving greater weights and farther pixels receiving smaller weights. The weight distribution follows a Gaussian distribution. During the filtering process, the Gaussian kernel size is set to five by five, with a standard deviation parameter of 1.0. The center of the kernel has the largest weight, the edges have the smallest weights, and the sum of all weights within the kernel equals one. The convolution operation slides the Gaussian kernel over the continuous intensity distribution map. At each location, the sum of the products of the kernel's pixels and the corresponding weights is calculated as the output pixel value. Specifically, the output value for a given pixel is the sum of the corresponding Gaussian weights for that pixel and its 24 surrounding pixels. This weighted average eliminates random noise in the image and smoothes the intensity distribution, resulting in a silo dust routing intensity map with excellent continuity and smoothness.

[0048] In a specific embodiment, the process of performing the two-dimensional spatial reconstruction of the device dust coupling coefficient matrix according to the capsule space coordinate mapping table may specifically include the following steps: Calculate the center coordinate position of each crusher dust feature capsule in the original image according to the capsule space coordinate mapping table to obtain the capsule center coordinate array; The coupling coefficient values ​​in the equipment dust coupling coefficient matrix are spatially matched according to the capsule center coordinate array to obtain a coordinate-intensity correspondence table; Based on the coordinate-intensity correspondence table, a two-dimensional array structure with the same size as the cement plant monitoring image is constructed to obtain the initialization intensity matrix; Fill the intensity values ​​in the coordinate-intensity correspondence table into the corresponding coordinate positions of the initialized intensity matrix to obtain a discrete intensity distribution map.

[0049] Specifically, the capsule spatial coordinate mapping table calculates the center coordinate position of each crusher dust feature capsule in the original image based on the convolution layer parameters of the capsule network and the index position of the capsule in the feature map. The calculation process uses a reverse mapping method to determine the corresponding position of the capsule in the 512 by 512 original image according to the row and column index of the capsule in the 252 by 252 feature map. In the specific calculation, let the row index of a capsule in the feature map be r and the column index be c. Since the primary capsule layer uses a convolution operation with a stride of 2, the horizontal coordinate of the original image center coordinate corresponding to this capsule is equal to the column index c multiplied by the stride of 2 plus the convolution kernel radius of 4, and the vertical coordinate is equal to the row index r multiplied by the stride of 2 plus the convolution kernel radius of 4. For example, the original image center coordinate of the capsule in the 50th row and 80th column of the feature map is 80 times 2 plus 4, which is equal to 164, and the vertical coordinate is 50 times 2 plus 4, which is equal to 108. This calculation process is performed on 63,504 crusher dust feature capsules one by one, generating a capsule center coordinate array containing horizontal and vertical coordinate pairs. The array length is 63,504, and each element contains the original image coordinate information of the corresponding capsule.

[0050] The coupling coefficient values ​​in the equipment dust coupling coefficient matrix are spatially matched to the capsule center coordinate array to establish a direct correspondence between capsule strength and spatial position. The matching process pairs each capsule's coupling coefficient value with its corresponding spatial coordinate, forming a triplet data structure consisting of abscissa, ordinate, and intensity value. The equipment dust coupling coefficient matrix is ​​a 63,504 by 10 two-dimensional array, where each row corresponds to a crusher dust signature capsule and each column corresponds to the coupling strength of a dust type capsule. The matching process selects the column corresponding to a specific dust type as the intensity value. Specifically, the coupling coefficient value for the kth capsule is extracted from the kth row of the equipment dust coupling coefficient matrix and paired with the coordinate value of the kth element in the capsule center coordinate array to form a coordinate-intensity triplet. For example, if the center coordinates of the 1000th capsule are abscissa 164 and ordinate 108, and the corresponding coupling coefficient value is 0.75, the triplet is abscissa 164, ordinate 108, and intensity 0.75. After matching each of the 63,504 capsules, a coordinate-intensity correspondence table is generated.

[0051] A coordinate-intensity correspondence table is constructed using a two-dimensional array structure of the same size as the cement plant monitoring image, creating a data container for storing spatially distributed intensity information. This construction process allocates memory for a two-dimensional array with 512 rows and 512 columns, with each element in the array corresponding to the intensity value of a pixel in the original image. Initialization sets all elements of the two-dimensional array to zero, indicating that no intensity information is present at any location. The row index of the array corresponds to the image's ordinate, and the column index corresponds to the image's abscissa. This combination of row and column indices allows for locating any pixel in the image. The data type of the initialized intensity matrix is ​​set to floating-point to support the storage of decimal intensity values. The matrix size is identical to that of the original cement plant monitoring image, ensuring accurate spatial correspondence. This matrix structure provides a standardized storage framework for subsequent intensity value filling operations, with each matrix element having a clear spatial location.

[0052] The intensity values ​​in the coordinate-intensity correspondence table are inserted into the corresponding coordinate positions of the initialization intensity matrix to complete the conversion from capsule intensity to image space intensity. The filling process traverses each triple in the coordinate-intensity correspondence table, extracting its horizontal coordinate, vertical coordinate, and intensity value information, and then assigning the intensity value to the element at the corresponding position in the initialization intensity matrix. Specifically, if the coordinates of a triple are horizontal coordinate x and vertical coordinate y, and the intensity value is v, then an assignment operation is performed to set the value of the element in the initialization intensity matrix at row y and column x to v. Because the number of capsules (63,504) is less than the total number of image pixels (262,144), the filled matrix exhibits a sparse distribution. Some locations contain non-zero intensity values, corresponding to dust features detected by the capsules, while most locations remain zero, indicating areas not covered by the capsules. After filling, a discrete intensity distribution map is obtained, which visually displays the spatial distribution pattern of the crusher dust features in grayscale format. High-intensity areas correspond to locations with high dust concentrations, while zero-value areas correspond to locations without dust detection or capsule coverage.

[0053] In a specific embodiment, step S4 includes: Count the total area and average intensity value of the area above the intensity threshold in the silo dust route intensity map to obtain the dust area statistical parameters; The dust concentration index of the cement plant is obtained by multiplying the dust area statistical parameters and the calibration coefficient. The cement plant dust concentration index is input into the multivariate linear regression model for concentration numerical conversion to obtain the quantitative dust concentration value; The quantitative dust concentration values ​​are classified based on the preset concentration level classification standards to obtain the dust concentration level assessment results of the production line.

[0054] Specifically, statistical processing of areas above the intensity threshold in the silo dust routing intensity map utilizes a combination of pixel traversal and conditional judgment to extract spatial and intensity information about valid dust areas. The statistical process sets the intensity threshold at 60% of the map's maximum intensity value. The threshold value is calculated after scanning the entire map to find the maximum intensity value. A pixel traversal operation examines the intensity value of each pixel in the 512x512 map row by row and column by column. When the pixel intensity exceeds the set threshold, the pixel is marked as a valid dust pixel and its coordinates and intensity information are recorded. The total area is calculated by summing the number of valid dust pixels, with each pixel representing a unit area. The total number of valid pixels represents the total area of ​​the dust region. The average intensity value is calculated by summing the intensity values ​​of all valid dust pixels and dividing it by the total number of valid pixels to obtain the average intensity level of the dust region. For this calculation, assuming the total number of valid dust pixels is N, with intensity values ​​ranging from I1, I2, to IN, the average intensity value is equal to the sum of all intensity values ​​divided by the total number of pixels, N. These two parameters constitute the dust area statistical parameters, where the total area reflects the dust diffusion range and the average intensity value reflects the dust concentration level.

[0055] The statistical parameters of the dust area are multiplied by a calibration coefficient to convert the statistically derived spatial and intensity information into a comprehensive concentration index. The calibration coefficient is obtained through simultaneous measurements in an actual cement plant environment. Real concentration data is collected at the same monitoring points using standard equipment such as a laser dust detector, establishing a correspondence between route intensity and actual concentration. The calibration process includes two parameters: the area calibration coefficient and the intensity calibration coefficient. The area calibration coefficient converts the number of pixels into actual physical area, while the intensity calibration coefficient converts route intensity values ​​into concentration contributions. The multiplication formula is: the cement plant dust concentration index equals the total area multiplied by the area calibration coefficient, the average intensity value, and the intensity calibration coefficient. This multi-parameter product calculation comprehensively considers the spatial and intensity distribution characteristics of dust. In the calculation, let the total area be A pixels, the area calibration coefficient be Ka, the average intensity value be I, and the intensity calibration coefficient be Ki. The concentration index is equal to the product of A times Ka times I times Ki. This calculation method compresses two-dimensional route intensity information into a one-dimensional concentration index, facilitating subsequent quantitative processing.

[0056] The cement plant dust concentration index is input into a multiple linear regression model for concentration numerical conversion, establishing a precise mapping relationship from the comprehensive index to the specific concentration value. The multiple linear regression model is trained based on a large amount of historical data, and the input variables include multiple characteristic parameters such as the cement plant dust concentration index, the maximum intensity value of the mapping diagram, and the aspect ratio of the effective dust area. The mathematical form of the regression model is that the quantitative dust concentration value is equal to the constant term plus the concentration index multiplied by the first regression coefficient, the maximum intensity value multiplied by the second regression coefficient, and the aspect ratio multiplied by the third regression coefficient. The model training process uses the least squares method to optimize the regression coefficients, and the optimal parameter combination is determined by minimizing the squared error between the predicted value and the measured value. The concentration numerical conversion process substitutes the input multiple characteristic parameters into the trained regression equation to calculate the quantitative dust concentration value in milligrams per cubic meter. In the conversion calculation, let the concentration index be CRI, the maximum intensity value be Imax, the aspect ratio be R, the corresponding regression coefficients be β1, β2, β3, and the constant term be β0. Then the quantitative concentration value is equal to β0 plus β1 multiplied by CRI plus β2 multiplied by Imax plus β3 multiplied by R.

[0057] Preset concentration classification standards classify quantitative dust concentration values, converting continuous concentration values ​​into discrete categories to facilitate environmental management and decision-making. These classification standards are based on environmental protection authorities' air quality standards and the cement industry's emission requirements, and include four levels: mild, moderate, severe, and serious. The classification process uses a threshold comparison method, setting the concentration range boundaries for each level and determining the level of concentration through numerical comparison. Specifically, 0 to 50 mg / m³ is classified as mild, 50 to 150 mg / m³ as moderate, 150 to 300 mg / m³ as severe, and 300 mg / m³ and above as severe. The classification algorithm uses conditional statements to compare the quantitative concentration values ​​against each level threshold. When the concentration value falls within a certain range, the corresponding level indicator is output. The classification results are presented as text to assess the production line's dust concentration level, providing a visual representation of the pollution severity in the monitored area.

[0058] Figure 2 This is a schematic diagram of the dust concentration level evaluation results of the production line of the cement plant dust intelligent identification method based on image processing in the embodiment of this application. Figure 2As shown, the figure shows the change trend of the quantitative dust concentration value obtained by calculating the cement plant dust concentration index and processing with the multivariate linear regression model during the 24-hour continuous monitoring of the cement plant. The horizontal axis in the figure represents the monitoring time (hours), and the vertical axis represents the dust concentration value after concentration quantification processing (mg / cubic meter). The background shaded areas correspond to the four concentration levels set by the present invention: mild level (0-50 mg / cubic meter), medium level (50-150 mg / cubic meter), severe level (150-300 mg / cubic meter) and severe level (300+ mg / cubic meter). The two key data points of 285 mg / cubic meter and 450 mg / cubic meter marked in the figure correspond to the severe and severe dust pollution states mentioned in the embodiment, respectively.

[0059] In a specific embodiment, step S5 includes: The connected domain analysis algorithm is used to process the silo dust routing intensity map to obtain a set of connected domains of independent dust areas. Calculate the geometric characteristic parameters of each connected domain in the connected domain set of independent dust regions to obtain the dust source geometric characteristic database; The dust source geometric feature database is input into the support vector machine classifier to perform equipment type identification processing and obtain the dust source equipment classification result; According to the spatial layout information of cement plant equipment, the dust source equipment classification results are processed by position matching and association to obtain the cement equipment associated dust source identification results.

[0060] Specifically, the silo dust routing intensity map is processed using a connected domain analysis algorithm, employing an eight-connected neighborhood search method to identify interconnected dust regions. The connected domain analysis algorithm sets an intensity threshold and converts the map into a binary image. Pixels above the threshold are marked as foreground pixels with a value of 1, and pixels below the threshold are marked as background pixels with a value of 0. The eight-connected neighborhood search process scans the binary image row by row, starting from the upper left corner. When a foreground pixel is encountered, a depth-first search or breadth-first search algorithm is initiated, searching the eight adjacent locations of that pixel (including the top, bottom, left, right, and four diagonal directions), classifying all connected foreground pixels as belonging to the same connected domain. During the search, a label matrix is ​​used to record the connected domain number of each pixel to avoid repeated searches for labeled pixels. Once a connected domain search is completed, the image is scanned again for the next unlabeled foreground pixel until the entire image is scanned. Area filtering removes small connected domains with fewer than 200 pixels to avoid noise interference, resulting in a set of connected domains for independent dust regions. Each connected domain represents a spatially independent dust source region.

[0061] The geometric characteristic parameters of each connected domain in the set of connected domains of independent dust regions are calculated, including four key indicators: centroid coordinates, bounding rectangle, aspect ratio, and compactness. The centroid coordinates are calculated by taking the weighted average of the coordinates of all pixels within the connected domain. Assuming that the connected domain contains N pixels with coordinates from x1y1 to xNyN, the centroid's horizontal coordinate is equal to the sum of all pixel horizontal coordinates divided by the total number of pixels, and the centroid's vertical coordinate is equal to the sum of all pixel vertical coordinates divided by the total number of pixels. The bounding rectangle calculation finds the minimum and maximum values ​​of the horizontal coordinates in the connected domain to determine the left and right boundaries of the rectangle, while the minimum and maximum values ​​of the vertical coordinates determine the upper and lower boundaries of the rectangle. The rectangle width is equal to the maximum horizontal coordinate minus the minimum horizontal coordinate plus 1, and the rectangle height is equal to the maximum vertical coordinate minus the minimum vertical coordinate plus 1. The aspect ratio is calculated as the ratio of the length to the width of the bounding rectangle. When the aspect ratio is greater than 1, the connected domain has a strip-like distribution, while when the aspect ratio is close to 1, the connected domain has a circular or square distribution. Compactness is calculated using the formula: 4 times pi multiplied by the area of ​​the connected domain divided by the square of the perimeter. A compactness close to 1 indicates a nearly circular connected domain, while a lower compactness indicates an irregular or dispersed connected domain. These four geometric characteristic parameters constitute a dust source geometric characteristic database, laying the foundation for subsequent equipment type identification.

[0062] A database of dust source geometric features is fed into a support vector machine classifier for equipment type identification. This classifier uses a pre-trained classification model to distinguish the dust characteristics generated by different cement equipment. During the training phase, the support vector machine classifier uses a large amount of labeled data to learn the geometric characteristic patterns of dust from different equipment. Crusher dust typically exhibits high compactness and a moderate aspect ratio, conveyor belt dust exhibits a strip-like distribution with a high aspect ratio, and silo dust exhibits a diffuse distribution with low compactness. The classification process inputs the four-dimensional feature vector of each connected domain into the trained support vector machine model. The model uses a kernel function to map the feature vector to a high-dimensional space and constructs an optimal separating hyperplane within this space to distinguish different classes. The decision function calculates the distance from the input feature vector to the separating hyperplane and determines the classification result and confidence level based on the positive and negative sign of the distance and its absolute value. During the classification process, the support vector machine outputs one of three categories for each connected domain: crusher type, conveyor belt type, or silo type, along with a classification confidence score. A high-confidence classification result indicates a close match between the feature pattern and the training data, resulting in the dust source equipment classification result.

[0063] The cement plant's equipment spatial layout information is used to perform position matching and association processing on the dust source equipment classification results, establishing a precise correspondence between dust sources and specific equipment. The equipment spatial layout information contains detailed information such as the precise coordinates, equipment type, and equipment number of all major equipment within the cement plant. This information is obtained from factory design drawings or field measurements. The position matching algorithm calculates the Euclidean distance between the coordinates of the centroid of each connected domain and the location of each device. The distance calculation formula is the square root of the square of the difference in the abscissa plus the square of the difference in the ordinate between the two points. The matching process selects devices of the same classification type as the connected domain and then searches for the device closest to the centroid of the connected domain as a matching target. A distance threshold of 50 pixels is set; matches exceeding this distance are considered unreliable and marked as unmatched. The association process associates successfully matched connected domains with the corresponding equipment, recording associated information such as equipment number, equipment type, dust intensity, and pollution level. This results in a cement equipment-associated dust source identification result, which details the dust emission status and pollution contribution of each device.

[0064] In a specific embodiment, the process of calculating the geometric characteristic parameters of each connected domain in the set of connected domains of independent dust regions may specifically include the following steps: Calculate the centroid coordinates and circumscribed rectangle size of each connected domain in the set of connected domains of independent dust regions to obtain the characteristic parameters of the crusher dust source position; The aspect ratio and regional compactness values ​​are calculated based on the location characteristic parameters of the crusher dust source to obtain the shape characteristic parameters of the conveyor belt dust source. Based on the shape characteristic parameters of the conveyor belt dust source and the area of ​​the connected domain, the characteristic vector combination processing is performed to obtain the comprehensive characteristic vector of the silo dust source; The comprehensive feature vectors of the silo dust source are processed according to the connected domain number to construct a feature database to obtain the dust source geometric feature database.

[0065] Specifically, the centroid coordinates of each connected domain in the set of connected domains within the independent dust region are calculated using a weighted average of pixel coordinates to determine the geometric center of the connected domain. The centroid coordinate calculation process traverses all pixels within the connected domain, extracts the abscissa and ordinate values ​​of each pixel, and then calculates the arithmetic mean of these coordinates. In the calculation, assume that the connected domain contains N pixels with pixel coordinates ranging from x1y1, x2y2, and so on to xNyN. The centroid abscissa is equal to the sum of the abscissas of all pixels divided by the total number of pixels, N, and the centroid ordinate is equal to the sum of the ordinates of all pixels divided by the total number of pixels, N. The bounding rectangle size is calculated by finding the extreme values ​​of the pixel coordinates within the connected domain and determining the minimum rectangle that encloses the entire connected domain. The algorithm traverses all pixels within the connected domain, recording the minimum xmin and maximum xmax of the abscissa, and the minimum ymin and maximum ymax of the ordinate. The bounding rectangle's width is equal to xmax minus xmin plus 1, and its height is equal to ymax minus ymin plus 1. The coordinates of the upper left corner of the rectangle are xmin and ymin, and the coordinates of the lower right corner are xmax and ymax. The center-of-gravity coordinates and the size of the circumscribed rectangle constitute the position characteristic parameters of the crusher dust source. The center-of-gravity coordinates represent the spatial position of the dust source in the image, and the size of the circumscribed rectangle represents the spatial distribution range of the dust source. These parameters directly reflect the spatial characteristics of dust generated by cement plant equipment.

[0066] Crusher dust source location characteristic parameters are calculated using aspect ratio and regional compactness to numerically quantify the shape characteristics of connected domains. The aspect ratio is calculated by dividing the length of the bounding rectangle by its width to obtain the shape ratio parameter. When the bounding rectangle is taller than the width, the aspect ratio is equal to the height divided by the width; when the width is wider than the height, the aspect ratio is equal to the width divided by the height, ensuring that the aspect ratio is always greater than or equal to 1. An aspect ratio close to 1 indicates that the connected domain is close to a square or circle, while an aspect ratio far greater than 1 indicates that the connected domain exhibits a strip-like or elliptical distribution. Regional compactness is calculated using the circularity formula to measure the regularity of the connected domain's shape. This is calculated as 4 times pi multiplied by the area of ​​the connected domain divided by the square of the perimeter. In the compactness calculation, the area of ​​the connected domain is equal to the total number of pixels it contains, and the perimeter is calculated by counting boundary pixels: pixels with at least one neighboring pixel that does not belong to the connected domain. The compactness value ranges from 0 to 1, with values ​​closer to 1 indicating that the connected domain shape is closer to a perfect circle, and smaller values ​​indicating a more irregular or scattered shape. The aspect ratio and area compactness constitute the shape characteristic parameters of the conveyor belt dust source, which can effectively distinguish the geometric morphological characteristics of dust generated by different equipment.

[0067] The shape characteristic parameters of conveyor belt dust sources and the area of ​​connected domains are combined into a feature vector to construct a multidimensional feature descriptor for dust source type identification. The feature vector combination process combines the three numerical parameters—aspect ratio, regional compactness, and connected domain area—into a three-dimensional vector. The first element of the vector is the aspect ratio, the second is the regional compactness, and the third is the connected domain area. The connected domain area is directly equal to the total number of pixels contained in the connected domain and reflects the spatial extent of dust diffusion. Before feature vector combination, numerical normalization is required to bring feature parameters of different dimensions and numerical ranges to the same scale to prevent a single feature parameter from dominating the entire vector due to excessively large values. This normalization method uses maximum-minimum normalization: the minimum value of each feature parameter across all connected domains is subtracted, and then divided by the difference between the maximum and minimum values. This normalized value is then distributed between 0 and 1. The three normalized parameters are combined to form a comprehensive feature vector for the silo dust source. This vector comprehensively describes the location, shape, and scale of the connected domain, providing standardized input data for subsequent machine learning classification algorithms.

[0068] The comprehensive feature vectors of silo dust sources are processed into a feature database based on connected domain numbers, establishing a structured feature storage and indexing system. During the feature database construction process, each connected domain is assigned a unique identifier. The numbers are assigned incrementally based on the order in which they are discovered in the image, starting from 1 and progressing until all connected domains are numbered. The database table structure includes fields such as the connected domain number, the horizontal coordinate of the centroid, the vertical coordinate of the centroid, the width and height of the bounding rectangle, the aspect ratio, the region compactness, the connected domain area, the normalized aspect ratio, the normalized compactness, and the normalized area. During data storage, all feature parameters of each connected domain are stored in the database table according to the corresponding field names, forming structured data records. The index construction process creates a primary key index for the connected domain number, a spatial index for the centroid coordinate, and a numerical index for the feature parameters, facilitating subsequent fast query and retrieval operations. The dust source geometric feature database not only stores the feature information of individual connected domains but also contains spatial relationships and statistical information between connected domains, such as global statistical parameters such as the total number of connected domains, average area, and area variance, forming a dust source feature description system.

[0069] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A cement plant dust intelligent identification method based on image processing, characterized in that: The method comprises: Step S1: feature extraction and processing of cement plant monitoring images are performed through a capsule network architecture to obtain a crusher dust feature capsule group; Step S2: performing routing connection processing on the crusher dust feature capsule group and the conveyor belt dust capsule layer according to a dynamic routing algorithm to obtain an equipment dust coupling coefficient matrix; Step S3: spatially reconstructing the equipment dust coupling coefficient matrix to obtain a silo dust routing intensity map; Step S4: performing concentration quantification processing on the silo dust route intensity map by calculating the cement plant dust concentration index to obtain a production line dust concentration level assessment result; In step S5, a multi-source separation process is performed on the silo dust routing intensity map according to a connected domain analysis algorithm to obtain a cement equipment-associated dust source identification result.

2. The method for intelligent identification of cement plant dust based on image processing according to claim 1 is characterized in that: The S1 step includes: Perform histogram equalization and gamma correction on the cement plant monitoring image to obtain a standardized pre-processed image; Inputting the standardized preprocessed image into the primary capsule layer of the capsule network architecture for convolution kernel feature extraction processing to obtain a crusher region capsule vector group; Performing nonlinear transformation processing on the crusher region capsule vector group based on the ReLU activation function to obtain an activated crusher feature vector; The activated crusher feature vector is subjected to length restriction processing according to a vector length normalization algorithm to obtain a crusher dust feature capsule group.

3. The method for intelligent identification of cement plant dust based on image processing according to claim 1 is characterized in that: The S2 step comprises: The crusher dust feature capsule group is input as a primary capsule into the dynamic routing algorithm for initialization processing to obtain an initial coupling coefficient matrix; Iteratively updating the initial coupling coefficient matrix according to the consistency between the prediction vector and the output vector to obtain an updated coupling coefficient matrix; Performing weight normalization processing on the updated coupling coefficient matrix based on a softmax normalization function to obtain a normalized coupling weight matrix; Performing a weighted sum operation on the normalized coupling weight matrix and the conveyor belt dust capsule layer to obtain a capsule layer output vector; The coupling coefficient of the capsule layer output vector is determined according to the vector length calculation to obtain the equipment dust coupling coefficient matrix.

4. The method for intelligent identification of cement plant dust based on image processing according to claim 3 is characterized in that: The iterative updating process of the initial coupling coefficient matrix according to the consistency between the prediction vector and the output vector to obtain the updated coupling coefficient matrix includes: Calculating the prediction vector between the crusher dust feature capsule group and the conveyor belt dust capsule layer to obtain an inter-capsule prediction vector matrix; Performing a dot product operation on the inter-capsule prediction vector matrix and the capsule layer output vector to obtain a consistency score matrix; Performing weighted accumulation processing on the consistency score matrix and the initial coupling coefficient matrix to obtain an accumulated coupling coefficient matrix; The accumulated coupling coefficient matrix is ​​updated by three rounds of iterative cycles to obtain an updated coupling coefficient matrix.

5. The method for intelligent identification of cement plant dust based on image processing according to claim 1 is characterized in that: The S3 step includes: Extracting the receptive field position information of each crusher dust feature capsule in the equipment dust coupling coefficient matrix to obtain a capsule space coordinate mapping table; Performing a two-dimensional spatial reconstruction process on the device dust coupling coefficient matrix according to the capsule space coordinate mapping table to obtain a discrete intensity distribution map; Performing continuous expansion processing on the discrete intensity distribution map based on a bilinear interpolation algorithm to obtain a continuous intensity distribution map; The continuous intensity distribution map is subjected to Gaussian filtering and smoothing processing to obtain a silo dust routing intensity map.

6. The method for intelligent identification of cement plant dust based on image processing according to claim 5 is characterized in that: The two-dimensional spatial reconstruction processing of the device dust coupling coefficient matrix is ​​performed according to the capsule space coordinate mapping table to obtain a discrete intensity distribution map, including: Calculate the center coordinate position of each crusher dust feature capsule in the original image according to the capsule space coordinate mapping table to obtain a capsule center coordinate array; Performing spatial position matching processing on the coupling coefficient values ​​in the equipment dust coupling coefficient matrix according to the capsule center coordinate array to obtain a coordinate-intensity correspondence table; Based on the coordinate-intensity correspondence table, a two-dimensional array structure of the same size as the cement plant monitoring image is constructed to obtain an initialized intensity matrix; The intensity values ​​in the coordinate-intensity correspondence table are filled into the corresponding coordinate positions of the initialized intensity matrix to obtain a discrete intensity distribution map.

7. The method for intelligent identification of cement plant dust based on image processing according to claim 1 is characterized in that: The S4 step comprises: Counting the total area and average intensity value of the area above the intensity threshold in the dust route intensity map of the silo to obtain dust area statistical parameters; Performing a product operation on the dust area statistical parameters and the calibration coefficient to obtain a cement plant dust concentration index; Inputting the cement plant dust concentration index into a multiple linear regression model for concentration numerical conversion processing to obtain a quantitative dust concentration value; The quantified dust concentration value is classified based on a preset concentration level classification standard to obtain a production line dust concentration level assessment result.

8. The method for intelligent identification of cement plant dust based on image processing according to claim 1 is characterized in that: The step S5 comprises: Performing a connected domain analysis algorithm on the silo dust routing intensity map to obtain a connected domain set of independent dust regions; Calculating the geometric characteristic parameters of each connected domain in the set of connected domains of the independent dust regions to obtain a dust source geometric characteristic database; Inputting the dust source geometric feature database into a support vector machine classifier to perform equipment type identification processing to obtain a dust source equipment classification result; The dust source equipment classification results are subjected to position matching and association processing according to the cement plant equipment spatial layout information to obtain cement equipment associated dust source identification results.

9. The method for intelligent identification of cement plant dust based on image processing according to claim 8, characterized in that: The step of calculating the geometric characteristic parameters of each connected domain in the set of connected domains of the independent dust regions to obtain a dust source geometric characteristic database includes: Calculating the coordinates of the center of gravity and the size of the circumscribed rectangle of each connected domain in the set of connected domains of the independent dust regions to obtain characteristic parameters of the crusher dust source position; Calculate the aspect ratio and regional compactness value according to the crusher dust source position characteristic parameters to obtain the conveyor belt dust source shape characteristic parameters; Performing feature vector combination processing based on the shape characteristic parameters of the conveyor belt dust source and the area of ​​the connected domain to obtain a comprehensive feature vector of the silo dust source; The comprehensive feature vector of the silo dust source is processed according to the connected domain number to construct a feature database to obtain a dust source geometric feature database.

Images